AI Policy · Daily

Commerce moved to shut a year-old workaround that had let firms headquartered in China acquire Nvidia and AMD advanced chips through offshore subsidiaries. Anthropic agreed to extend its Mythos vulnerability-finding model to the EU Agency for Cybersecurity (ENISA) under its Project Glasswing partner program. Rival super PACs Public First Action and Leading the Future, backed by Anthropic and OpenAI respectively, have pledged more than $100 million and are running ads against each other's 2026 midterm picks. NIST rebranded its AI Safety Institute Consortium as the NIST AI Consortium, reorganizing the roughly 280-member body around six task groups centered on measurement, innovation and adoption.

I.AI Policy Today

Commerce closes year-old loophole letting Chinese firms buy Nvidia and AMD chips through overseas subsidiaries

The U.S. Department of Commerce posted guidance Sunday saying it will enforce license requirements for Nvidia Rubin, Blackwell and AMD MI350x advanced chips sold to entities headquartered in China, even when those entities are located outside China, including in Malaysia. The opening was created in May 2025, when Commerce announced it would not enforce the AI Diffusion rule the Biden administration issued in its final days. One chip industry source familiar with supply chains estimated that hundreds of thousands of chips had been exported through the loophole, per Reuters. The new guidance does not require data centers to stop using chips already in place or to cut off servicing of advanced computing items.

Read at SCMP ↗ • Read at Yahoo Finance ↗

Anthropic to grant EU cybersecurity agency ENISA access to Mythos vulnerability-finding model

Anthropic plans to give the European Union Agency for Cybersecurity, known as ENISA, access to its Mythos model, Bloomberg reported Monday. Mythos is Anthropic's restricted frontier model for finding and exploiting computer vulnerabilities. Since April it has been previewed to 12 partner organizations including Amazon, Apple, Cisco, CrowdStrike and Palo Alto Networks under Anthropic's Project Glasswing initiative. The arrangement comes four days after a senior EU official told CNBC that Brussels wanted to "intensify" U.S. talks on access to advanced cyber AI models.

Read at Bloomberg ↗ • Read at TechCrunch ↗

Anthropic-aligned and OpenAI-aligned super PACs target each other's candidates as 2026 AI spending tops $100 million

Public First Action, linked to Anthropic, and Leading the Future, the OpenAI-aligned super PAC, have spent roughly $24 million on the 2026 cycle and pledged more than $100 million in combined commitments, The New York Times reported. The two groups have refused to coordinate on shared candidates and are running opposing ads in some races. Leading the Future has raised $140 million from Trump-aligned tech leaders including OpenAI co-founder Greg Brockman, the venture firm Andreessen Horowitz and Palantir co-founder Joe Lonsdale, and has spent $750,000 backing Rep. Andy Barr in the Kentucky Senate race plus $1.1 million on Georgia House primaries, per Axios. Public First has raised $30 million from individual donors plus a $20 million Anthropic contribution earmarked for public education spending, and put $2.3 million into supporting RAISE Act author Alex Bores in NY-12, per The Washington Post. In North Carolina's 4th District, Public First ran a $1.6 million February ad blitz backing Rep. Valerie Foushee, who held off challenger Nida Allam by roughly 1,200 votes.

Read at NYT ↗ • Read at Washington Post ↗ • Read at Axios ↗

NIST renames AI Safety Institute Consortium to NIST AI Consortium and reopens member applications under innovation framing

The National Institute of Standards and Technology announced Friday that it has renamed its AI Safety Institute Consortium, known as AISIC, to the NIST AI Consortium. The agency restructured the work into six task groups focused on AI measurement, innovation and adoption, and opened a public letters-of-interest call for new member organizations. NIST founded the consortium in 2023, bringing together more than 280 organizations. The relaunch follows NIST's earlier rebranding of its AI Safety Institute as the Center for AI Standards and Innovation.

Read at NIST ↗

Check Point finds 2026 midterm cyber threats target campaign systems and AI misinformation rather than voting machines

Cybersecurity firm Check Point Software Technologies published research Monday concluding that threats to the 2026 elections target email accounts, websites and fundraising platforms used by campaigns, donors and voters, rather than voting machines or ballot counting systems, CyberScoop reported. Email carries 82% of malicious attacks against these groups, per the report. AI amplified misinformation joins phishing and impersonation as a major content threat, per The Hill. Check Point also counted roughly 9,500 stolen ActBlue passwords and 6,500 stolen WinRed passwords circulating since the cycle began, plus about 1,300 newly registered websites containing the word "election" and 4,010 containing the word "vote" registered in January alone.

Read at The Hill ↗ • Read at CyberScoop ↗

II.China Watch

BYD launches chip for self-driving vehicles in domestic silicon push

BYD founder Wang Chuanfu unveiled the Xuanji A3 at the company's launch event last Thursday, describing it as China's first 4nm autonomous driving chip with support for Level 3 and Level 4 driving capabilities, TechNode reported. The chip has entered mass production, and a three-chip configuration delivers combined computing power that BYD says is enough to support L3 and L4 functions in production vehicles. BYD has built its own semiconductor business since 2002, and the Xuanji A3 is the first chip from that line to reach the 4nm-class node. The launch puts a Chinese automaker into volume production with a domestically designed 4nm-class chip, narrowing Chinese carmakers' exposure to U.S. export controls on advanced automotive compute supplied from abroad.

Read at TechNode ↗

MiniMax begins A-share IPO process after Hong Kong rally

MiniMax's board resolved on May 31 to explore issuing renminbi shares for a Shanghai STAR Market listing, after the AI model developer signed a pre-IPO tutoring agreement with CITIC Securities filed to China's securities regulator on May 29, Caixin reported. MiniMax's Hong Kong shares have risen more than 400% since their Jan. 9 listing, putting market capitalization near HK$260 billion ($33 billion) as of May 29. The move would create a dual A-share and H-share listing structure. The filing would give Chinese regulators a domestic public equity vehicle for a frontier AI model maker at a moment when U.S. outbound investment screening has constrained Chinese AI firms' access to American capital.

Read at Caixin ↗

Unitree Robotics tees up Shanghai listing as China's first "embodied intelligence" stock

Humanoid robot maker Unitree Technology was set for a June 1 listing committee review at the Shanghai Stock Exchange, positioning the company to become China's first publicly traded "embodied intelligence" stock, Pandaily reported. Unitree's business spans high-performance humanoid robots, quadruped robots, robotic components and embodied AI models. The company has been one of the more visible Chinese players in the embodied AI race, with general-purpose humanoid platforms widely deployed in research and industrial demos. A successful listing would establish a domestic public equity benchmark for China's humanoid robotics push, a sector Chinese authorities have designated as strategic and recently brought under a Beijing-based lifecycle management platform.

Read at Pandaily ↗

X-Square Robot open-sources world model built around robot action events

Chinese embodied AI firm X-Square Robot released WALL-WM as open source on May 29, describing it as the world's first event-level prediction embodied AI world model, Pandaily reported. The model organizes supervision and training data around action-grounded semantic events rather than fixed time windows, which the firm said lets robot models train against task objectives rather than fixed pixel sequences. The release follows an earlier WALL-family open release from the same company. Chinese central planners have designated humanoid robotics a strategic industrial priority, putting open source frontier work in the sector in line with Beijing's industrial policy.

Read at Pandaily ↗

III.Industry & Market Watch

SoftBank pledges up to €75 billion ($87 billion) for 5 gigawatts of French AI data center capacity

SoftBank Group plans to invest up to €75 billion ($87 billion) to build 5 gigawatts of AI data center capacity in France, with a first phase of €45 billion ($52 billion) targeting 3.1 gigawatts in the Hauts-de-France region by 2031, the Financial Times reported. SoftBank announced the pledge in a Saturday statement. Founder Masayoshi Son met President Emmanuel Macron at the Élysée Palace on Monday during the Choose France Summit, where the deal was folded into Macron's €93 billion ($108 billion) foreign investment package. Initial sites will be in Dunkirk, Bosquel and Bouchain, with Schneider Electric a partner at Dunkirk, per Fortune. The commitment follows SoftBank's March plan to channel up to $500 billion into 10 gigawatts of capacity in Ohio.

Read at FT ↗ • Read at Fortune ↗

Foundation Future Industries targets U.S. military front-line humanoid deployment within 18 months with Eric Trump as adviser

Foundation Future Industries, a San Francisco startup founded in 2024, plans to deploy its humanoid robots with the U.S. military within 12 to 18 months after testing Phantom MK-1 units for combat zone logistics in Ukraine, CNBC reported. The company holds $24 million in research contracts across the Army, Navy and Air Force for inspection, logistics and weapons handling, and is preparing improved Phantom 2 units with double the 44-pound payload of Phantom 1 for Ukraine deployments this year. Eric Trump joined the company as chief strategy adviser after being an investor; Democratic Sen. Elizabeth Warren called the contracts "corruption in plain sight." CEO Sankaet Pathak said some weaponized uses of the Phantom robots will retain human confirmation in the decision loop, while in certain time-critical scenarios the robots will need to make fully autonomous decisions, per The Next Web.

Read at CNBC ↗ • Read at The Next Web ↗

U.S. Army's Operation Jailbreak puts more than 50 defense vendors at Fort Carson to make legacy weapons systems interoperable

Operation Jailbreak, an Army-led hackathon at Fort Carson, Colorado, has drawn hundreds of engineers from more than 50 defense companies to expose proprietary interfaces and let missile systems, tanks, drones and other Army equipment exchange data in real time, DefenseScoop reported. Initial participants included Anduril, Boeing, General Dynamics, L3Harris, Lockheed Martin, Northrop Grumman, Palantir, Perennial Autonomy and RTX, with Army Secretary Dan Driscoll calling the effort the "largest hackathon in human history." Driscoll said the initiative was modeled on Ukraine's Delta battlefield network and was accelerated by drone defense demands from the Iran War, with some "jailbroken" systems already deployed to the Middle East and a goal of pushing most updates to U.S. Central Command within 30 days. The effort sits inside the Army's Right to Integrate program and is paired with a $65 billion budget shift to fund the Army Transformation Initiative, which Driscoll said supports venture capital style funding and revised acquisition processes.

Read at DefenseScoop ↗ • Read at FT ↗

IV.Global & Geopolitics

UK Ministry of Defence weighs allowing autonomous AI lethal strikes in narrowly defined cases

UK military officials are pushing an internal policy change that would allow AI systems to make lethal strike decisions without human approval in "exceptional circumstances," the Financial Times reported on May 30. The proposed carve-out would create a doctrinal exception to current UK military rules requiring human approval for autonomous lethal action. The reporting diverges from the human decision-making position Vice President JD Vance laid out at the U.S. Air Force Academy last week, when he said human decision-makers, not AI, must control lethal force in war.

Read at FT ↗

UK Home Office contracts for AI facial age estimation on disputed asylum seekers as more than 100 children's groups protest

The UK Home Office awarded a contract Friday, May 29, to deploy AI facial age estimation on asylum seekers whose age is contested, the Guardian reported. A coalition of more than 100 refugee children's organizations is opposing the rollout, with the Refugee and Migrant Children's Consortium set to publish an opposition report in June. The groups say trauma, malnutrition and stress can alter the facial markers the model uses to estimate age.

Read at Guardian ↗

New Zealand Parliament passes automated decision-making welfare bill to modernise benefit services

New Zealand's Social Security (Modernisation) Amendment Bill, which enables broader use of automated decision-making across the welfare system, passed its third reading in Parliament, Inside Government NZ reported. Social Development and Employment Minister Louise Upston said the bill implements a Budget 2025 change and does not affect eligibility settings, with automation applied only to straightforward decisions that do not require discretion. The system covers rules-based automation rather than generative AI such as ChatGPT, which is excluded, RNZ reported. The bill strengthens safeguards requiring the Ministry of Social Development to manage bias, ensure transparency and maintain human oversight.

Read at Inside Government NZ ↗ • Read at RNZ ↗